Enhancing Emotion Recognition from Speech Signals Using Deep Learning and Open-Source Datasets

In this work, we focus on emotions from speech signals using deep learning methods on the RAVDESS database. The audio signals are preprocessed by means of sampling, normalization and the subsequent feature extraction in the form of Mel Frequency Cepstral Coefficients (MFCCs). Classification of emotions is then performed using a Convolutional Neural Network often abbreviated as CNN. The model's performance is therefore measured with performance metrics including accuracy, precision, recall, and F1 score, to give a measure of the model's capabilities in detecting emotions from voices. The results of the confusion matrix analysis help to understand better the ability of the model to predict different emotions. The analysis proves the significance of deep learning approaches, for example, CNNs that significantly outperform classical approaches in terms of emotion recognition rates. The study also finds out the audio features, which have very high impact in the classification of emotions and contains finding that will be useful for further improvements of affective computing.